A well-known term in the context of binaural signal enhancement is the so-called ``cocktail-party effect,'' denoting that, with the aid of binaural hearing, humans can concentrate on one talker in the presence of competing ones. It has further been established that with binaural hearing a desired signal and noise can be separated more effectively than with monaural hearing. Binaural auditory models may help to simulate these capabilities by providing front-ends that allow for better separation of a mix of sound sources. In a specific Bochum version of a so-called ``cocktail-party'' processor, i.e., a processor to enhance speech in a ``cocktail-party'' situation, the binaural processor of the auditory model is used to control a Wiener filter [35,34]. This is accomplished by first identifying the position of a desired talker in space, and then estimating its S/N ratio with respect to competing talkers and other noise signals. The system performs its computation within critical-bands. In the case of two competing talkers, the desired signal can be recovered to reasonable intelligibility, even when its level is 15 dB lower than that of the competing one. Application possibilities for this kind of systems are numerous, such as tools for editing binaural recordings, front ends for signal-processing hearing aids, speech-recognizers and hands-free telephones. In general, binaural signal enhancement may be used to build better ``microphones'' for acoustically adverse conditions. As stated above, the cues provided by models of the subcortical auditory system, and contained in binaural-activity patterns, must consequently be evaluated in adequate ways. The next section deals with this problem.
Esprit Project 8579/MIAMI (Schomaker et al., '95)